Graph Neural Network Policies and Imitation Learning for Multi-Domain Task-Oriented Dialogues

TL;DR

This paper proposes using graph neural network-based structured policies combined with imitation learning to improve multi-domain task-oriented dialogue systems, addressing challenges of large state-action spaces and sparse rewards.

Contribution

It introduces a novel approach integrating graph neural networks with imitation learning for multi-domain dialogues, demonstrating improved performance over standard policies.

Findings

Structured policies outperform standard policies in multi-domain dialogues.

Graph neural networks effectively model domain relationships.

Imitation learning enhances policy effectiveness in complex dialogue scenarios.

Abstract

Task-oriented dialogue systems are designed to achieve specific goals while conversing with humans. In practice, they may have to handle simultaneously several domains and tasks. The dialogue manager must therefore be able to take into account domain changes and plan over different domains/tasks in order to deal with multidomain dialogues. However, learning with reinforcement in such context becomes difficult because the state-action dimension is larger while the reward signal remains scarce. Our experimental results suggest that structured policies based on graph neural networks combined with different degrees of imitation learning can effectively handle multi-domain dialogues. The reported experiments underline the benefit of structured policies over standard policies.